Moche-AI

Gallery

About

Moche-AI is a short-term-rental concierge SaaS centered on a per-property Property Brain™ and guest portal. It is built with Next.js, Supabase, Stripe, OpenRouter, and Firecrawl, with production deployments on Vercel; public go-live is planned for 2027, with pre-launch access and controlled pilots preceding it.How it works:The product connects host-managed property knowledge to a guest-facing concierge and a host operations dashboard. Property data stays scoped to its property; guest answers rely on approved knowledge, and unresolved requests can route to the host for review or action.Property knowledge: Hosts provide property facts and source materials. Direct host-entered values can be written to the Brain, while extracted or AI-generated updates remain reviewable proposals; guest retrieval rechecks approval, source text, property, visibility, and lifecycle before returning knowledge.Guest portal: Guests use a mobile web experience for property Q&A, local recommendations, Extras, service needs, multilingual interactions, and host messaging. The portal uses retrieval-grounded answers rather than general-purpose chat alone.Host operations: The dashboard links overview and notification actions with guest conversation and Extras-request workflows so hosts can manage guest follow-up.AI routing: Approved exact answers can return without a completion or embedding call. Routine concierge work uses a lighter route; complex concierge, Guest Needs, translation, and Brain operations use protected stronger routes with constrained fallback behavior.Delivery boundaries: Production deployments and offline validation do not establish live provider delivery. The September release separates stored-message success from Twilio acceptance, and the subsequent AI-workflow release left Twilio provider work paused pending a separate live acceptance pass.Architecture decisions:Property-scoped, host-approved knowledge as the answer source — retrieval checks current approval and scope; AI extraction and learned answers remain reviewable before becoming guest-facing facts. This protects property-specific accuracy and prevents unreviewed model output from becoming authoritative.Task-tiered routing with an exact-answer fast path — known approved answers bypass unnecessary model and embedding calls, while complex work uses stronger configured routes instead of silently degrading to a cheaper model.Separate pre-launch access from public launch — host-side access and controlled product pilots can precede the stated 2027 public go-live; a production deployment alone is not a decision to begin general commercial launch.